Robust Fast Inter-Bin Image Registration for Undersampled Coronary MRI Based on a Learned Motion Prior.

Area of Science:

• Medical Imaging
• Image Registration
• Artificial Intelligence

Background:

• Coronary Magnetic Resonance Angiography (CMRA) is crucial for cardiac assessment.
• Image artifacts from undersampling degrade CMRA quality and hinder motion estimation.
• Accurate motion compensation is vital for high-quality CMRA reconstruction.

Purpose of the Study:

• To develop a 3D nonrigid registration method for accurate displacement field estimation from artifact-corrupted CMRA.
• To address motion estimation biases caused by undersampling artifacts in CMRA.
• To improve the quality of motion-compensated reconstruction for highly accelerated 3D CMRA.

Main Methods:

• A novel registration framework, UNROLL, was developed using a 3D U-Net initializer and a deep unrolling network.
• A supervised learning framework with training labels from fully-sampled images was employed.
• The unrolling network learns a task-specific motion prior to reduce artifact-induced biases.

Main Results:

• UNROLL demonstrated improved accuracy in motion estimation and motion-compensated CMRA reconstruction compared to baseline methods at 6-fold acceleration.
• The method maintained superior displacement field accuracy even at 11-fold acceleration, exceeding training data.
• UNROLL achieved a rapid computational time of only 2 seconds for a whole 3D volume.

Conclusions:

• The proposed UNROLL method effectively incorporates a learned respiratory motion prior for accurate motion estimation in highly accelerated CMRA.
• UNROLL offers a fast and accurate solution for estimating displacement fields from low-quality CMRA images.
• This technique holds significant potential for enhancing motion-compensated reconstruction in accelerated 3D CMRA.